mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 10:02:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add some extra info to diagnostics

Browse Source
This commit is contained in:
Daniel Povey 2022-03-15 13:49:15 +08:00
parent 1e5455ba29
commit 21ebd356e7
1 changed files with 17 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
egs/librispeech/ASR/transducer_stateless/diagnostics.py
View File

@ -79,7 +79,7 @@ def get_diagnostics_for_dim(dim: int, tensors: List[Tensor],
dim: the dimension to analyze, with 0 <= dim < tensors[0].ndim dim: the dimension to analyze, with 0 <= dim < tensors[0].ndim
options: options object options: options object
sizes_same: true if all the tensor sizes are the same on this dimension sizes_same: true if all the tensor sizes are the same on this dimension
stats_type: either "abs" or "positive" or "eigs" or "value, stats_type: either "abs" or "positive" or "eigs" or "value",
imdictates the type of stats imdictates the type of stats
we accumulate, abs is mean absolute value, "positive" we accumulate, abs is mean absolute value, "positive"
is proportion of positive to nonnegative values, "eigs" is proportion of positive to nonnegative values, "eigs"
@ -129,12 +129,23 @@ def get_diagnostics_for_dim(dim: int, tensors: List[Tensor],
percentiles.append(stats[index].item()) percentiles.append(stats[index].item())
percentiles = [ '%.2g' % x for x in percentiles ] percentiles = [ '%.2g' % x for x in percentiles ]
percentiles = ' '.join(percentiles) percentiles = ' '.join(percentiles)
return f'percentiles: [{percentiles}]' ans = f'percentiles: [{percentiles}]'
else: else:
stats = stats.tolist() ans = stats.tolist()
stats = [ '%.2g' % x for x in stats ] ans = [ '%.2g' % x for x in ans ]
stats = '[' + ' '.join(stats) + ']' ans = '[' + ' '.join(ans) + ']'
return stats if stats_type == "value":
norm = (stats ** 2).sum().sqrt().item()
mean_abs = stats.abs().mean().item()
# This norm is useful because it is strictly less than the largest
# sqrt(eigenvalue) of the variance, which we print out, and shows,
# speaking in an approximate way, how much of that largest eigenvalue
# can be attributed to the mean of the distribution.
ans += f', norm={norm:.2g}, mean_abs={mean_abs:.2g}'
else:
mean = stats.mean().item()
ans += f', mean={mean:.2g}'
return ans